Subscribe Bookmark RSS Feed

Change-point analyzer

agneshb

Community Trekker

Joined:

Jan 19, 2014

When analyzing trends in data (i.e. process monitoring) is there a way in JMP to analyze when exactly a shift started?

Thank you,

Agnès

4 REPLIES
louv

Staff

Joined:

Jun 23, 2011

Have you checked out Analyze>Quality and Process>Control Chart>Multivariate Control Chart?

There is a change point detection capability there.

Control Charts in general have rules to alarm out of control situations as well.

Byron_JMP

Staff

Joined:

Apr 26, 2012

Agnès,

There are a couple of things that are going to make this easy vs. very hard.

     1. How big is the change?  <1std is hard

     2. How many points before and after the suspected change? 25-30 is great, less than that, combined with a small shift and the signal to noise ratio starts to get small.


There are lots of approaches to this problem. The easiest is to just use the Tests in the control chart platforms.

The next easiest approach is to use the Partition platform. Put the time variable in the X role and your measured response in the Y role. Interactively split the data to find segments of time with different means. With this approach, when I have lots of splits, its a good idea to be skeptical of segments that are short, and splits with small log worth score where the difference in the mean is small.


The multivariate control chart's change point analyzer is nice, but it can only find one change point. Then you have to manually partition the data and repeat the change point on each of the following partitions. 

agneshb

Community Trekker

Joined:

Jan 19, 2014

Thank for both answers, I'll try to use the partitioning platform.

Is there any way to run the change point detection on individual parameter charts (IR in my case) even if it's not in the menu?

Dan_Obermiller

Joined:

Apr 3, 2013

Agnes,

Rather than trying to figure out how to add change point detection to an IR chart (which would involve scripting and programming the math for the change point detection), I would recommend still using the Multivariate Control Chart. Just specify your single variable. This platform still works with only one variable.

Dan Obermiller